a multi-purpose ontology-based approach

7/31/2019 A Multi-Purpose Ontology-Based Approach

1/6

A Multi-Purpose Ontology-Based Approach

for Personalized Content Filtering and Retrieval

David Vallet, Ivn Cantador, Miriam Fernndez, and Pablo Castells

Universidad Autnoma de Madrid, Escuela Politcnica Superior

{david.vallet,ivan.cantador,miriam.fernandez,pablo.castells}@uam.es

Abstract

Personalized multimedia access aims at enhancing

the retrieval process by complementing explicit user

requests with implicit user preferences. We propose

and discuss the benefits of the introduction of ontolo-

gies for an enhanced representation of the relevantknowledge about the user, the context, and the domain

of discourse, as a means to enable improvements in the

retrieval process and the performance of adaptive ca-

pabilities. We develop our proposal by describing

techniques in several areas that exemplify the explota-

tion of the richness and power of formal and explicit

semantics descriptions, and the improvements therein.

1. Introduction

Personalized multimedia access aims at enhancing

the retrieval process by complementing explicit user

requests with implicit user preferences, to better meetindividual user needs [5]. Automatic user modeling

and personalization has been a thriving area of re-

search for nearly two decades, gaining significant pres-

ence in commercial applications around the mid-90s.

Personalizing a content retrieval system involves

considerable complexity, mainly because finding im-

plicit evidence of user needs and interests through their

behavior is not an easy task. This difficulty is often

considerably increased by an imprecise and vague rep-

resentation of the semantics involved in user actions

and system responses, which makes it even more diffi-

cult to properly pair user interests and content descrip-

tions. The ambiguity of terms used in this representa-tion, the unclear relationships between them, their het-

erogeneity, especially in current ever-growing large-

scale networked environments such as the WWW,

often constitute a major obstacle for achieving an accu-

rate personalization, e.g. when comparing user prefer-

ences to content items, or users among themselves.

In this paper we argue for the introduction of on-

tologies [8] as an enhanced representation of the rele-

vant knowledge about the domain of discourse, about

users, about contextual conditions, involved in the re-

trieval process, as a means to enable significant im-

provements in the performance of adaptive content

retrieval services. We exemplify our point by describ-ing the development of advanced features and en-

hancements in specific areas related to personalization

where the ontology-based approach shows its benefit,

including:

Basic personalized content search and browsingbased on user preferences.

Dynamic contextualization of user preferences.

Dynamic augmented social networking and collabo-rative filtering.

Domain ontologies and rich knowledge bases play a

key role in the models and techniques that we propose

in the above areas, as will be described in the sequel.

The approaches presented in this paper share and ex-ploit a common representation framework, thus obtain-

ing multiple benefits from a shared single ontology-

rooted grounding. Furthermore, it will be shown that

modular semantic processing strategies, such as infer-

ence, graph processing, or clustering, over networked

ontology concepts, may be reused and combined to

serve multiple purposes.

The rest of the paper is organized as follows. Sec-

tion 2 introduces the basic approach for the ontology-

oriented representation of semantic user preferences,

and its application to personalized content search and

retrieval. Following this, section 3 describes an ap-

proach for the dynamic contextualization of semanticuser preferences, and section 4 shows the extension of

the techniques described in previous sections to multi-

user environments, based on collaborative personaliza-

tion strategies. Finally, some conclusions are given in

section 5.

First International Workshop on Semantic Media Adaptation and Personalization (SMAP'06)0-7695-2692-6/06 $20.00 2006

Authorized licensed use limited to: IEEE Xplore. Downloaded on November 1, 2008 at 07:41 from IEEE Xplore. Restrictions apply.


2/6

2. Ontology-based personalization for con-

tent retrieval

Most personalized retrieval techniques (e.g. col-

laborative filtering) keep and process long records of

accessed documents by each user, in order to infer

potential preferences for new documents (e.g. by find-ing similarities between documents, or between users).

The data handled by these techniques have been rather

low-level and simple: document IDs, text keywords

and topic categories at most [9], [12]. The recent pro-

posals and achievements towards the enrichment of

multimedia content by formal, ontology-based, seman-

tic descriptions open new opportunities for improve-

ment in the personalization field from a new, richer

representational level [2], [5]. We see the introduction

of ontology-based technology in the area of personal-

ization as a promising research direction [7]. Ontolo-

gies enable the formalization of user preferences in a

common underlying, interoperable representation,whereby user interests can be matched to content

meaning at a higher level, suitable for conceptual rea-

soning.

An ontology-based representation is richer, more

precise, less ambiguous than a keyword-based model.

It provides an adequate grounding for the representa-

tion of coarse to fine-grained user interests (e.g. inter-

est for individual items such as a sports team, an actor,

a stock value) in a hierarchical way, and can be a key

enabler to deal with the subtleties of user preferences.

An ontology provides further formal, computer-

processable meaning on the concepts (who is coaching

a team, an actors filmography, financial data on astock), and makes it available for the personalization

system to take advantage of. Moreover, an ontology-

rooted vocabulary can be agreed and shared (or

mapped) between different systems, or different mod-

ules of the same system, and therefore user prefer-

ences, represented this way, can be more easily shared

by different players. For instance, a personalization

framework may share a domain ontology with a

knowledge-based content analysis tool that extracts

semantic metadata from a/v content, conforming to the

ontology [2]. On this basis, it is easier to build algo-

rithms that match preference to content, through the

common domain ontology.In an ontology-based approach, semantic user pref-

erences may be represented as a vector of weights

(numbers from 0 to 1), representing the intensity of the

user interest for each concept in a domain ontology [5].

If a content analysis tool identifies, for instance, a cat

in a picture, and the user is known to like cats, the per-

sonalization module can find how the user may like the

picture by comparing the metadata of the picture, and

the preferred concepts in the user profile. Based on

preference weights, measures of user interest for con-

tent units can be computed, with which it is possible to

discriminate, prioritize, filter and rank contents (a col-

lection, a catalog section, a search result) in a personal

way.Furthermore, ontology standards backed by interna-

tional consortiums (such as the W3C), and the corre-

sponding available processing tools, support inference

mechanisms that can be used to further enhance per-

sonalization, so that, for instance, a user interested in

animals (superclass of cat) is also recommended pic-

tures of cats. Inversely, a user interested in lizards,

snakes, and chameleons can be inferred to be inter-

ested in reptiles with a certain confidence. Also, a user

keen of Sicily can be assumed to like Palermo, through

the transitive locatedIn relation. In fact, it is even pos-

sible to express complex preferences based on generic

conditions, such as athletes that have won a goldmedal in the Olympic Games.

3. Contextual personalization

The shallowest consideration is sufficient to notice

that human preferences are complex, variable and het-

erogeneous, and that not all preferences are relevant in

every situation [13]. For instance, if a user is consis-

tently looking for some contents in the Formula 1 do-

main, it would not make much sense that the system

prioritizes some Formula 1 picture with a helicopter in

the background just because the user happens to have a

general interest for aircrafts. In other words, in thecontext ofFormula 1, aircrafts are out of (or at least far

from) context. Context is a difficult notion to grasp and

capture in a software system, and the elements than

can, and have been considered in the literature under

the notion of context are manifold: user tasks and

goals, computing platform, network conditions, social

environment, physical environment, location, time,

noise, external events, text around a word, visual con-

text of a graphic region, to mention a few.

Complementarily to the ones mentioned, we pro-

pose a particular notion, for its tractability and useful-

ness in semantic content retrieval: that ofsemantic

runtime context, which we define as the backgroundthemes under which user activities occur within a

given unit of time. Using this notion, a finer, qualita-

tive, context-sensitive activation of user preferences

can be defined. Instead of a uniform level of personal-

ization, user interests related to the context are priori-

tized, discarding the preferences that are out of focus.

The problems to be addressed include how to represent




3/6

such context and determine it at runtime, and how the

activation of user preferences should be related to it,

predicting the drift of user interests over time. Our

approach is based on a concept-oriented context repre-

sentation, and the definition of distance measures be-

tween context and preferences as the basis for the dy-

namic selection of relevant preferences [13].A runtime context is represented (is approximated)

in our approach as a set of weighted concepts from the

domain ontology. This set is obtained by collecting the

concepts that have been involved, directly or indi-

rectly, in the interaction of the user (e.g. issued queries

and accessed items) with the system during a retrieval

session. The context is built in such a way that the im-

portance of concepts fades away with time (number of

user requests back when the concept was referenced)

by a decay factor.

Once a context is thus built, the contextual activa-

tion of preferences achieved by a computation of the

semantic similarity between each user preference andthe set of concepts in the context. In spirit, the ap-

proach consists of finding semantic paths linking pref-

erences to context, where the considered paths are

made of existing semantic relations between concepts

in the domain ontology. The shorter, stronger, and

more numerous such connecting paths, the more in

contexta preference is considered. The proposed tech-

niques to find these paths use a form of Constraint

Spreading Activation (CSA) strategy [6]. In fact, in the

proposed approach a semantic expansion of both user

preferences and the context takes place, during which

the involved concepts are assigned preference weights

and contextual weights, which decay as the expansion

grows farther from the initial sets. This process can

also be understood as finding a sort of fuzzy semantic

intersection between user preferences and the semantic

runtime context, where the final computed weight of

each concepts represents the degree to which it belongs

to each set (see Figure 1).

Initial user preferences

Extended user preferences

Initial runtime context

Extended context

Domain

concepts

Semantic

user preferences

Semantic

runtime context

Contextualised

user preferences

Contextualised

user preferences

Figure 1. Contextual activation of semantic user pref-erences.

Finally, the perceived effect of contextualization is

that user interests that are out of focus, under a given

context, are disregarded, and only those that are in the

semantic scope of the ongoing user activity (the inter-

section of user preferences and runtime context) are

considered for personalization. The inclusion or exclu-

sion of preferences is in fact not binary, but ranges ona continuum scale, where the contextual weight of a

preference decreases monotonically with the semantic

distance between the preference and the context.

Contextualized preferences can be understood as an

improved, more precise, dynamic, and reliable

representation of user preferences, and as such they

can be used directly for the personalized ranking of

content items and search results, as described in the

previous section, or they can be input to any system

that exploits this information in other ways, such as the

one described in the next section.

4. Augmented social networking and col-laborative filtering

When the system perspective is widened to take in

contextual aspects of the user, it is often relevant to

consider that in most cases the user does not work in

isolation. Indeed, the proliferation of virtual communi-

ties, computer-supported social networks, and collec-

tive interaction (e.g. several users in front of a Set-

Top-Box), call for further research on group modeling,

opening new problems and complexities. A variety of

group-based personalization functionalities can be en-

abled by combining, comparing, or merging prefer-

ences from different users, where the expressive powerand inference capabilities supported by ontology-based

technologies can act as a fundamental piece towards

higher levels of abstraction [3], [4].

4.1. Semantic group profiling

Group profiling can be understood under the ex-

plicit presence of a-priori given user groups, or as an

activity that involves the automatic detection of im-

plicit links between users by the system, in order to put

users in contact with each other, or to help them bene-

fit from each others experience. In the first view, col-

laborative applications may be required to adapt togroups of people who interact with the system. These

groups may be quite heterogeneous, e.g. age, gender,

intelligence and personality influence on the percep-

tion and demands on system outputs that each member

of the groups may have. The question that arises is

how can the system adapt itself to the group in such a

way that each individual benefits from the results.




4/6

Figure 2. Group profiling by aggregation of individualuser profiles.

We have explored the combination of the ontology-

based profiles defined in section 2 to meet this pur-

pose, on a per concept basis, following differentstrategies from social choice theory [11] for combining

multiple individual preferences. In our approach, user

profiles are merged to form a shared group profile, so

that common content recommendations are generated

according to this new profile (see Figure 2). Our pre-

liminary experiments have shown that improved re-

sults can be obtained from the accuracy and expressiv-

ity of the ontology-based representation as proposed in

this approach [3].

4.2. Semantic social networking

Even when explicit groups are not defined, usersmay take advantage of the experience of other users

with common interests, without needing to know each

other. The issue of finding hidden links between users

based on the similarity of their preferences or historic

behavior is not a new idea. In fact, this is the essence

of the well-known collaborative recommender systems

[1], where items are recommended to a certain user

concerning those of her interests shared with other

users or according to opinions, comparatives, and rat-

ings of items given by similar users.

However, in typical approaches, the comparison be-

tween users and items is done globally, in such a way

that partial, but strong and useful similarities may be

missed. For instance, two people may have a highly

coincident taste in cinema, but a very divergent one insports. The opinions of these people on movies could

be highly valuable for each other, but risk to be ig-

nored by many collaborative recommender systems,

because the global similarity between the users might

be low.

In recommendation environments there is an under-

lying need to distinguish different layers within the

interests and preferences of the users. Depending on

the current context, only a specific subset of the seg-

ments (layers) of a user profile should be considered in

order to establish her similarities with other people

when a recommendation has to be performed. Models

of social networks partitioned into different commonsemantic layers can achieve more accurate and con-

text-sensitive results.

The definition and generation of such models can

be facilitated by a more accurate semantic description

of user preferences, as supported by ontologies. A

multi-layered approach to social networking can be

developed by dividing user profiles into clusters of

cohesive interests, so that several layers of social net-

works are found. This provides a richer model of in-

terpersonal links, which better represents the way peo-

ple find common interests in real life.

Taking advantage of the relations between con-

cepts, and the (weighted) preferences of users for the

concepts, we have defined a system that clusters the

semantic space based on the correlation of concepts

appearing in the preferences of individual users [4].

After this, user profiles are partitioned by projecting

the concept clusters into the set of preferences of each

user (see Figure 3). Then, users can be compared on

the basis of the resulting subsets of interests, in such a

way that several, rather than just one, (weighted) links

can be found between two users.

Figure 3. Multilayer generation of social links between users: 1) the initial sets of individual interests are expanded, 2)domain concepts are clustered based on the vector space of user preferences, and 3) users are clustered in order toidentify the closest class to each user.




5/6

Multilayered social networks are potentially useful

for many purposes. For instance, users may share pref-

erences, items, knowledge, and benefit from each

others experience in focused or specialized conceptual

areas, even if they have very different profiles as a

whole. Such semantic subareas need not be defined

manually, as they emerge automatically with our pro-posed method. Users may be recommended items or

direct contacts with other users for different aspects of

day-to-day life.

In addition to these possibilities, our two-way space

clustering, which finds clusters of users based on the

clusters of concepts found in a first pass, offers a rein-

forced partition of the user space that can be exploited

to build group profiles for sets of related users. These

group profiles enable efficient strategies for collabora-

tive recommendation in real-time, by using the merged

profiles as representatives of classes of users.

The degree of membership of the obtained subpro-

files to the clusters, and the similarities among them,can be used to define social links to be exploited by

collaborative filtering systems. We report early ex-

periments with real subjects in [4], where the emergent

social networks are applied to a variety of collabora-

tive filtering models, showing the feasibility of the

clustering strategy.

4.3. Semantic profile expansion for collabora-

tive group profiling

In real scenarios, user profiles tend to be very scat-

tered, especially in those applications where user profiles

have to be manually defined. Users are usually not will-ing to spend time describing their detailed preferences to

the system, even less to assign weights to them, espe-

cially if they do not have a clear understanding of the

effects and results of this input. On the other hand, ap-

plications where an automatic preference learning algo-

rithm is applied tend to recognize the main characteris-

tics of user preferences, thus yielding profiles that may

entail a lack of expressivity. To overcome this problem,

the semantic preference spreading mechanism described

in section 3 has proved highly useful for improving our

group profiling techniques as well.

Previous experiments without the semantic spreading

feature showed considerably poorer results. The profileswere very simple and the matching between the prefer-

ences of different users was low. Typically, the basic

user profiles provide a good representative sample of

user preferences, but do not reflect the real extent of user

interests, which results in low overlaps between the

preferences of different users. Therefore, the extension is

not only important for the performance of individual

personalization, but is essential for the clustering strat-

egy described in the previous subsection.

In very open collaborative environments, it is also

the case that not only direct evidence of user interests

needs to be properly completed in their semantic con-

text, but that they are not directly comparable with the

input from other users in its initial form. If the envi-ronment is very heterogeneous, the potential disparity

of vocabularies and syntax used by different users or

subsystems pose an additional barrier for collaborative

techniques. One of the major purposes for which on-

tologies are conceived is that of reflecting or achieving

a consensus between different parties in a common

knowledge space [8]. Therefore, they provide special-

purpose facilities to ensure the required interoperabil-

ity between semantic user spaces, and match descrip-

tions that are syntactically different but semantically

related.

5. Conclusion

In this paper we propose the introduction of ontolo-

gies as a key tool for moving beyond current state of

the art in the area of personalization. We show ways in

which ontology-driven representations can be used to

improve the effectiveness of different personalization

techniques, and we describe a set of functionalities

where ontologies are essential to achieve qualitative

enhancements. In our approach, ontologies are used to

model the domain of discourse in terms of which user

interests, content meaning, retrieval context, and social

relationships, can be described and analyzed.

The advantages of ontology-driven representations(expressiveness and precision, formal properties, infer-

ence capabilities, interoperability) enable further de-

velopments that exploit such capabilities, beyond the

ones proposed here, on top of the basic personalization

framework described in this paper.

A trade-off of our proposals is the cost and diffi-

culty of building well-defined ontologies and populat-

ing large-scale knowledge bases, which is not ad-

dressed in this paper. Recent research on these areas is

yielding promising results [10], in a way that any ad-

vancement on these problems can be played to the

benefit of our proposed achievements.

6. Acknowledgements

The research leading to this document has received

funding from the European Communitys Sixth Frame-

work (FP6-027685 MESH), and the Spanish

Ministry of Science and Education (TIN2005-06885).

However, it reflects only the authors views, and the




6/6

European Community is not liable for any use that may

be made of the information contained therein.

7. References

[1] M. Balabanovic and Y. Shoham, Content-Based Col-

laborative Recommendation, Communications of theACM40(3), March 1997, pp. 66-72.

[2] S. Bloehdorn, K. Petridis, C. Saathoff, N. Simou, V.Tzouvaras, Y. Avrithis, S. Handschuh, Y. Kompatsiaris,

S. Staab, M. G. Strintzis, Semantic Annotation of Im-

ages and Videos for Multimedia, 2nd European Seman-

tic Web Conference (ESWC 2005). Springer Verlag Lec-

ture Notes in Computer Science, Vol. 3532, 2005.

[3] I. Cantador, P. Castells, and D. Vallet, EnrichingGroup Profiles with Ontologies for Knowledge-Driven

Collaborative Content Retrieval, 1st International

Workshop on Semantic Technologies in Collaborative

Applications (STICA 2006), at the 15th IEEE Interna-

tional Workshops on Enabling Technologies (WETICE2006), Manchester, UK, June 2006.

[4] I. Cantador and P. Castells, Multilayered SemanticSocial Network Modeling by Ontology-Based User Pro-

files Clustering: Application to Collaborative Filtering,

15th International Conference on Knowledge Engineer-

ing and Knowledge Management (EKAW 2006), Po-

debrady, Czech Republic, October 2006, Springer Ver-

lag Lecture Notes in Computer Science, to appear.

[5] P. Castells, M. Fernndez, D. Vallet, P. Mylonas, andY. Avrithis, Self-Tuning Personalized Information Re-

trieval in an Ontology-Based Framework, 1st Interna-

tional Workshop on Web Semantics (SWWS 2005), AgiaNapa, Cyprus, July 2005. Springer Verlag Lecture

Notes in Computer Science, Vol. 3762, pp. 977-986.

[6] F. Crestani, Application of Spreading ActivationTechniques in Information Retrieval, Artificial Intelli-

gence Review 11, 1997, pp. 453-482.

[7] S. Gauch, J. Chaffee, and A. Pretschner, Ontology-based personalized search and browsing, WIAS Journal

1(3-4), 2003, pp. 219-234.

[8] T. R. Gruber, A Translation Approach to Portable On-tology Specification, Knowledge Acquisition 5, 1993,

pp. 199-220.

[9] G. Jeh, J. Widom, Scaling Personalized Web Search,International World Wide Web Conference (WWW

2003), Budapest, Hungary, 2003, pp. 271-279.

[10]A. Kiryakov, B. Popov, I. Terziev, D. Manov, D. Ogn-yanoff, Semantic Annotation, Indexing, and Retrieval,

Journal of Web Sematics 2(1), Elsevier, 2004, pp. 47-

49.

[11]J. Masthoff, Group Modeling: Selecting a Sequence ofTelevision Items to Suit a Group of Viewers, User

Modeling and User-Adapted Interaction 14(1), 2004,

pp. 37-85.

[12]A. Micarelli, F. Sciarrone, Anatomy and EmpiricalEvaluation of an Adaptive Web-Based Information Fil-

tering System, User Modelling and User-Adapted In-

teraction 14(2-3), 2004, pp. 159-200.

[13]D. Vallet, M. Fernndez, P. Castells, P. Mylonas, andY. Avrithis, Personalized Information Retrieval in

Context, 3rd International Workshop on Modeling and

Retrieval of Context (MRC 2006) at the 21st National

Conference on Artificial Intelligence (AAAI 2006), Bos-

ton, USA, July 2006.


a multi-purpose ontology-based approach

Documents